FTD MINDER.
NSCI + RPI IDEA institute · Live R Shiny app, four years and counting
Performance engineering and visualization work on FTD MINDER, the public dashboard that supports Bowles et al. (Cell, 2021) on frontotemporal dementia organoids. I shipped the Seurat sketch frame layer that made 30GB of single-cell RNA-seq data interactive on commodity hardware, and built several of the cell-type visualizations.
Role
1 of 3 engineers: performance + viz layer
Stack
R · Shiny · Seurat
Result
30GB → laptop-interactive, seconds-range loads
Status
Live since 2021 · supports Bowles et al. (Cell, 2021)
Reference: Bowles, K.R. et al., 'ELAVL4, splicing and glutamatergic dysfunction precede neuron loss in MAPT-mutation cerebral organoids,' Cell, July 26, 2021. doi.org/10.1016/j.cell.2021.07.003
30GB dataset
needs an HPC cluster
Seurat sketch frames
representative cells
Laptop-interactive
seconds-range loads
The problem at hand
Researchers working on frontotemporal dementia (FTD) at the Neural Stem Cell Institute generated 30GB of single-cell RNA-seq data from MAPT-mutation cerebral organoids. The dataset was load-bearing for the science (it underpins what became a 2021 Cell paper) and load-bearing for the field (collaborators across institutions needed to explore it independently). Manipulating it interactively required parallel CPU processing across multiple cores. Researchers without HPC access couldn't open the data at all.
The end goal: a public dashboard a researcher anywhere with a browser could open and explore.
The team was three students at RPI's IDEA institute working on the visualization layer. The brief was 'build a dashboard for the NSCI dataset.' Within that, my specific corner became performance engineering. The dashboard couldn't exist as a useful tool unless the underlying data structure supported real-time exploration on commodity hardware. The compression problem had to be solved before the visualization design mattered.
What I built
Seurat sketch frame layer (the load-bearing piece)
Sketch frames preserve representative cell populations from the full dataset, allowing real-time UMAP rendering and gene-expression queries without paging through 30GB of cells. It brought interactive load times into the seconds range. This is what made the dashboard publicly viable.
Several of the cell-type visualizations
UMAP embeddings annotated by cell type and MAPT genotype, plus related views. I collaborated with the other two students on the broader R Shiny app architecture and the rest of the visualization layer.
What I did not do: drive the science (that was the NSCI team), lead the overall app architecture (that was a team effort), or build every visualization on the dashboard. The honest framing is one of three engineers, owning the performance-engineering layer plus a portion of the visualization work.
The messy middle
From parallel processing to Seurat sketch frames
The first version of the dashboard ran against the full dataset using parallel processing across multiple CPU cores. That worked for the team but it was the wrong shape for the field: anyone without HPC access was locked out. Swapping in the sketch frame layer is the moment it became usable as a shared research tool, not an internal RPI artifact.
Outcome
The dashboard launched alongside the Cell paper in 2021 and remains live on RPI's research infrastructure as of 2026. The performance work took the dataset from researchers at HPC-equipped institutions to anyone with a browser. The Cell paper itself carries the science citation. The dashboard carries the engineering contribution.